A Screenwriter, A Horse, & An Oyster Walk into a Bar...

Whirlwind week of travel from one coast to the other— but the news doesn’t stop!
Although video content was a little light on the channel this week, on the plus side, we get to hang out in your Inbox!

I’ve covered a few AI-powered screenwriting platforms in the past, Plotdot being one of them. And while I feel most of them are pretty good for exploration and story beat ideas, I have yet to see anything one-shot an Oscar-winning screenplay…and I think most of you would agree.

So, when Google DeepMind's Fabula was demoed this week, naturally, there were a lot of hot takes on how Google is rolling in to kill screenwriters.

But as usual, hot takes are gonna come in hot and not get things right.

Fabula is a Gemini-powered research prototype aimed squarely at fiction writers, screenwriters, and playwrights. It's also not meant as a tool to replace writers, nor is it actually that new.

Fabula is the direct descendant of Dramatron, which the same DeepMind team (led by researcher Piotr Mirowski) published on arXiv back in 2022. So while most of our Deepmind eyes have been focused on Gemini, Nanobanana, Veo, etc— this team has quietly been working the narrative structure problem for about four years.

Architecturally, it's a two-layer editor: a top layer for the story plan (overall arc, scenes, beats), and a bottom layer for the actual script (dialogue, action, description). You can regenerate or revise at either level independently, and the structural skeleton persists through rewrites — so when you rework a beat, your long-form consistency doesn't go out the window. If you’ve tried to write longform with an LLM, that’s one of the biggest hurdles.

And while it can generate with consistency and history, again, this is something that is meant for the writer to remain in the driver's seat.

My Take: At the end of the day, Is Fabula is essentially a Google/Gemini version of Save the Cat?

Maybe, but here's an interesting twist: Fabula being grounded in classical narratology, and appears to be trained on stories drawn from a much more diverse set of cultures and narrative traditions, means — and this is the optimistic read — in the right hands, this tool could actually lead to more interesting stories, not fewer. A working writer using Fabula as a sidekick might end up pulling from narrative traditions they'd never have cracked open on their own. Think of stuff that Joseph Campbell DIDN’T cover in The Power Of Myth. That's a potentially pretty exciting outcome, and one I don't hear many people talking about.

The WGA Beat: One thing briefly worth flagging — the WGA and AMPTP actually just reached a tentative four-year agreement in early April, avoiding a strike and entering the member ratification phase. Their post-strike contract specifically limits how studios can use AI-generated literary material. A Google-branded script-development tool emerging publicly right now is either very well-timed or a very interesting coincidence. I'll let you decide which.

That said, I think it's important to remember that Hollywood does not own screenplays. And the WGA does not own stories. There are countless non-guild productions and screenplays written and produced every day.

The Third Act Twist: This is actually not in the wild as of yet. I have signed up to be an early tester, so I’ll report on it as soon as I can. In the meantime, if I hear more about availability, I’ll let you know.
Google has been pretty good about shipping lately, so keep an eye on Google Labs as well!

Happy Horse 1.0 — the upcoming video model from Alibaba's Future Life Lab, sitting inside the newly-formed Alibaba Token Hub (ATH) business unit — popped on Arena.ai this week with a pretty loud splash, briefly locking in #1 in both Text-to-Video and Image-to-Video (no-audio) leaderboards inside of 24 hours. As votes have accumulated since, the model has settled into a solid #2 slot, just behind Seedance 2.0 — but even that initial spike was enough to turn a lot of heads.

Here's the interesting bit, though: The technical lead on Happy Horse is Zhang Di, who until recently was VP at Kuaishou, where he architected Kling 1.0 and 2.0.

Yeah: Daddy Kling!!

So, he left Kuaishou in the fall of 2025, joined Alibaba in November, and four months later shipped a model that immediately made it to the top of the Arena leaderboards on debut. That’s pretty baller.

My Take: I'll be honest — I'm skeptical of the “Seedance Killer” claims but hopeful of a "Seedance 2.0 competitor," although even that is a bold claim right now, and we haven't actually seen that much from Happy Horse yet beyond the leaderboard numbers. Arena benchmarks can always be gamed, and until a model is in wide circulation, I tend to side eye most rankings.

Where You Can Actually Try It: For now, Arena.ai seems to be the best bet — and even that's a bit of a crapshoot. If you're lucky enough to get Happy Horse in battle mode, you'll get a generation you can actually eyeball against the competition. Roll the dice a few times and you may run across it, at which point we can finally start really judging for ourselves whether this thing is a real Seedance 2.0 competitor or not.

The Bigger Picture: I mean, this should come as no surprise: Word on the street is that basically every AI lab working on video is committing enormous GPU resources to training up a model that equals, if not beats, Seedance 2.0. It probably won't be too long before we start seeing Seedance's massive lead begin to thin out. Happy Horse may or may not turn out to be the model that closes that gap, but there is no doubt the race is on.

Same team, different vibe. Alibaba's ATH unit recently dropped Happy Oyster — their real-time interactive world model. Unlike Happy Horse, this one isn't a video generator in the usual sense, but more of a Genie-3 type World Simulator.

We’ve seen a few of these popping up lately, both as major lab products and on the open-source scene, but Happy Oyster’s model looks good enough that Alibaba's stock jumped 5.6% on the announcement, which does mean investors are watching this space.

And it has some interesting features beyond Genie 3.

It ships in two modes:

Both modes output synchronized audio + video.

My Take: Happy Oyster is, at its core, another iteration on Google's Genie 3 — and Genie 3 is, let's be clear, staggeringly impressive technology. But it was also released late last year and in AI time, basically makes it ancient. The other really honest part: I don't think most of us have really figured out the killer use case for these open-world, continuously-generating world simulators yet.

And that's fine! Honestly, I've never thought of Genie 3 — or Happy Oyster, or any of these real-time world models — as the end goal. What we're actually looking at right now are building blocks for whatever the next thing ends up being. Steps to the holodeck, perhaps. A new kind of production tool for virtual cinematography five years from now. Something we haven't fully imagined yet.

In the meantime, I'm always interested in checking out the latest iteration of this technology — and Happy Oyster is available for beta testing right now. You can sign up for the beta here.

And hey, about a month before we find out if Genie-4 is in the wings!

Props to @simeonnz for solving what has been a significant limitation with Seedance 2.0. If you've tried uploading audio files directly to Seedance, you’ve run into the issue where the model changes the lyrics, and sometimes even the song itself. I think they frame is as “inspiration”— but, well it can be pretty frustrating.

Simeon's workaround is really clean: save out a blank video with your song attached, and then upload that video as a Video Omni reference. That's it. You get perfect lip-sync and no style drifting on the audio.

Quick note on use case: The best case for using a straight audio file at all — instead of this video-omni workaround — is really only if you want to keep a character's voice consistent across generations. For music videos specifically, Simeon's approach is one to put in your back pocket!

Bonus Tip — Midjourney 6.1 for Face Recognition: While we're in Seedance-land — through a bunch of experimentation this week, I'm finding that Midjourney version 6.1 characters seem to have the most success getting past Seedance's face recognition blocker. Mileage may vary, as always. Interestingly, reposing or setting up shots in Nano Banana Pro tends to break it. So this usually only really works in the Omni model with a 6.1 character reference and additional location/prop references.

I’ve now done a bunch of tests and can say: 6.1 seems to be the one has the highest success rate.

The testing continues…as always…

As mentioned at the top: had a bit of a whirlwind travel week heading out to Mountain View, California and back within the span of two days. The gang at Utopai invited me out for the launch of their latest update.

It was a lot of fun being a fly on the wall for the chaos that is launch day. Spent a lot of time talking to leadership and the devs about the platform — and in particular, they were very keen on understanding our workflows and how they can be integrated into a future update.

I had some really insightful conversations with the research team (shoutout to Yinan & Hui!) who were kind enough to entertain and explain all my bonehead questions about model architectures and other math things that are WAY above my pay grade.

I spent most of the day there playing with the new interface, telling them where I was running into friction, and showing them how most of us tend to work — even giving them a BTS of the Dragon Blue project…which, honestly, was pretty much the same walkthrough in my video.

Spent a good amount of time talking to a famous animation director, who— I don’t think I’m allowed to say a name— but, I can tell you: There is no one on the planet who hasn’t seen one of his movies.

Some other creators were also in attendance, and it was a treat meeting AI Video School and Charles Curran — and while all three of us approach AI Filmmaking from different perspectives, there are a lot of core mechanics in iteration and revision that we all stressed we'd like to see more of. Again, I always appreciate that the team was clearly listening.

Rounded out that evening grabbing dinner with Jessie from Pika. If you've been around the AI Video space since the start, you'll no doubt remember the heyday that was the eternally long Pika 1.0 Beta — back when we all still generated on Discord…And Jessie was the "voice" of that channel! So, after all these years, it was great to have a chance to hang out!

My big takeaway for the week: We all work in and create synthetic realities, communicate through comments/WhatsApp groups, Discords, email chains, and Zoom/Meet calls — but the best days are when we can sit in the same room together and exchange ideas and just hang out.
Bonus if there is Pizza and cocktails!

Screwing my head back on for what I’m sure is going to be an action packed week! As Always I thank you for Reading…My name is Tim!